|Publication number||US7743023 B2|
|Application number||US 11/275,870|
|Publication date||Jun 22, 2010|
|Priority date||Feb 1, 2006|
|Also published as||CN101336435A, CN101336435B, EP1979845A1, EP1979845A4, US20070185932, WO2007089354A1|
|Publication number||11275870, 275870, US 7743023 B2, US 7743023B2, US-B2-7743023, US7743023 B2, US7743023B2|
|Inventors||Dan Teodosiu, David C. Steere, Muthukaruppan Annamalai, Nikolaj S Bjorner, Michael J Bohlander, Huisheng Liu, Patrick E Bozeman|
|Original Assignee||Microsoft Corporation|
|Export Citation||BiBTeX, EndNote, RefMan|
|Patent Citations (28), Non-Patent Citations (4), Referenced by (11), Classifications (8), Legal Events (3)|
|External Links: USPTO, USPTO Assignment, Espacenet|
Multi-master file replication systems have been built for many years using various approaches, such as state-based or event-based replication. Typically, these systems will synchronize a portion of the file system between two or more computing devices (nodes), such that changes made to files or directories on one node will be propagated to the other nodes.
Many times, the computing devices on which replication occurs tend to be computers that are not permanently accessible, either because the computing devices may not be running or may be disconnected from the Internet. Accordingly, in these environments, users can typically only access their replicated files when they have access to their computing device. Hence, the user who travels a great deal may experience great difficulties in accessing their files.
Various embodiments introduce the notion of a replication entity which implements a highly scalable file replication system. In one embodiment, the replication entity resides in the form of a scalable replication service. In at least some embodiments, the replication service provides a “drive in the sky” facility that can be used by individuals, such as subscribers, to synchronize their individual machines, such that their files are automatically replicated to a safe and always-on location. Alternatively or additionally, individuals such as subscribers can also access their files via a web-based interface when they are away from their machines.
The various embodiments described below introduce the notion of a replication entity which implements a highly scalable file replication system. The replication entity can reside in any suitable form such as in a service or replication hub or hubs.
In the example used in this document, the replication entity resides in the form of a scalable replication service. In at least some embodiments, the replication service provides a “drive in the sky” facility that can be used by individuals, such as subscribers, to synchronize their individual machines, such that their files are automatically replicated to a safe and always-on location. Alternatively or additionally, individuals such as subscribers can also access their files via a web-based interface when they are away from their machines.
The inventive embodiments address problems associated with scaling replication services to a very large number of subscribers. For example, simply scaling files system volumes to tens or hundreds of petabytes (PB) of storage poses significant cost, scalability, management, and reliability challenges, as will be appreciated by the skilled artisan. Stores that offer ideal characteristics in terms of scalability, management, and reliability can be built, but they will typically not have full file system semantics, due to the complexities associated with providing such semantics at that scale. Further, file system replication systems typically keep some replication metadata on the side. Since this metadata is updated separately from the file system update path, it can (and often does) get out of sync with the actual file system contents. Should that be the case, replicators will typically rescan the entire file system to rebuild an authoritative copy of the metadata. A rescan (especially if it occurs relatively frequently) can be a very onerous operation in a scalable service.
Accordingly, in at least some of the embodiments described below, these and other issues are addressed through a design that does not use a file system in the service. Instead, the design is based on using a scalable, self-managing, and reliable binary large object (blob) store (also called a “scalable file store” below) that does not provide full file system semantics. For instance, not providing full file system semantics may mean not supporting one or more of the following: a hierarchical file namespace, logical file names, file update operations, file truncation or extension, sparse files, file locking, or a filesystem change journal. When the “sky” node receives a file from one of the subscriber nodes, the file is received as a binary large object (blob) containing both actual file data and meta-data. Instead of unpacking that blob and storing it in a file system, however, the sky node simply stores it into the scalable blob store.
Additionally, in at least some embodiments, the replication metadata is merged with or stored alongside the scalable file store metadata. Thus, any updates to the blob store (whether originating from replication with one of the subscriber machines, or due to changes made via the web interface) can transact both the blob store metadata and the replication metadata at the same time. This mitigates synchronization issues between the replication metadata and the blob store and can thus obviate the need for rescans.
To provide some context for the discussion of scalable file replication below, consider the following on state-based replication. It is to be appreciated and understood that while the inventive embodiments are described in the context of state-based replication, the techniques described herein can also be applied to other types of replicators, such as event-based replicators.
File replicators are used to synchronize the content of file systems between two or more machines (computing devices), or “nodes”. Any modifications performed to the file system on one of the nodes will be propagated to the other nodes. Replicators are designed to deal with all possible race conditions, such as conflicting updates to the same file on multiple nodes and the like. The node file systems can be any of the usual ones, such as NTFS, FAT, XFS, etc.
File system modification events are processed locally on each node and accumulated in that node's replication metadata database 106. In the case of NTFS, the file system modification events can be read out of the file system's change journal 108, which records all file creations, deletions, and modifications. Other file systems may have different notification mechanisms. The node database's ID table 110 records the replication metadata for each of the replicated files and directories. This metadata is needed in addition to the actual file system data to efficiently determine what files need to be synchronized between two nodes and to identify and resolve conflicts and file moves.
In this example, upstream node 102 serves replicated content and the downstream node 104 receives replicated content. The vv-up and vv-down arrows signify communication between the nodes, such as the exchange of version vectors, which is designed to ensure that a minimum amount of updates can flow from the upstream node 102 to the downstream node 104. The space and updates arrows signify a credit system whereby the downstream node 102 can control the number of updates it receives from the upstream node 104 following a version vector exchange.
Downstream node 104 processes the updates using meet component 112 and decides, depending on the nature of the updates to the file, whether to download or delete files, as will be appreciated by the skilled artisan. When node 104 decides to download a file, it retrieves the file contents from upstream node 102 and places the file contents in a staging area 114. Files are exchanged between nodes using a get and put protocol.
The other components in this picture consist of the file system interface (FS) and the database 106 that contains metadata on what is on the FS. The database also contains identifiers that are global to all replicated nodes.
With regard to ID Table 110, consider now, in Table 1 just below, an overview of the main replication metadata that is stored in the ID Table 110.
Machine GUID of unique identifier
Version of unique identifier
Machine GUID of GVSN
Version of GVSN
Machine GUID of parent UID
Version of parent UID
Clock value (used to resolve file
Creation time. (used to resolve
Resource name. It is a variable
The first two fields (UidMachine and UidVersion) give the unique global identifiers for a resource. The unique global identifier does not change when a resource is updated—it is invariant under the lifetime of the resource.
Contrarily, the global version sequence numbers (i.e. GvsrMachine and GvsnVersion) change whenever a given resource is updated. The GvsnMachine field identifies the partner that made the change, and the GvsnVersion field identifies a timestamp of the change.
The ParentMachine and ParentVersion fields identify the unique identifier of the parent directory. Specifically, since files are being replicated in a file system, there is an ancestral relationship where the objects were replicated. The Parent fields are the fields used to identify a directory in which a file resides.
The Clock field provides an update time of a resource and is used to resolve update conflicts. Specifically, if there are two concurrent updates, the updates will typically have different clocks which are then used to resolve conflicts.
The CreateTime field is also used to resolve directory conflicts. The creation time of a fixed resource does not change.
The Name field is the name of a resource as it appears in the file system. This is not the fully qualified name, but rather the name relative to a parent.
Periodically, as shown in
The synchronization core is comprised of a handshake between two partners: the upstream and downstream nodes. As will be appreciated by the skilled artisan, synchronization always occurs from the upstream node to the downstream node. For bi-directional synchronization, two handshakes are needed. The two partners start the handshake by exchanging version vectors, which are a summary of the GVSN's of all of their files, as reflected in their respective ID tables. Based on the version vector received from its downstream partner, the upstream node determines what information (i.e. files) it has that its downstream partner has not yet seen. The upstream node then starts sending updates (i.e. information about files that downstream has not yet seen), and downstream node downloads content based on these updates.
As indicated in
Upon receiving the meta-data for a remote resource, a node must determine whether the file contents associated with the resource and its particular version is admissible. In particular, a file must pass the following filters:
An update that passes the above tests may still be conflicting, in the sense that it was made independently of the file version stored locally. This scenario is common when collaborators edit the same document at the same time on two different nodes. However, either version may be installed on a file system, so the update conflict does not violate any file system constraints.
Having discussed file replication, consider now scalable file storage.
Scalable File Storage
Scalable file storage systems have been built for many years using various approaches, including namespace partitioning (e.g. AFS, DFS), metadata and blob decoupling (e.g. Google's GFS), and clustered file systems (e.g. Sistina's GFS). The techniques and approaches described in this document are applicable to scalable file storage that decouples the metadata and blobs, although the techniques described herein could also be applied to other types of scalable file storage.
The scalable file system metadata is stored on one or more nodes 504 optimized for metadata storage and indexing. The metadata is replicated for availability and reliability. A SQL database is commonly used as the metadata store.
The data streams for the files are typically stored on one or more nodes 506 optimized for blob storage, as will be appreciated by the skilled artisan. Scalable blob stores generally provide reduced storage semantics as compared to a traditional file system, typically only providing object put, get, and delete operations. Objects written to the blob store are replicated for availability and reliability.
Clients perform scalable file system operations by calling remote HTTP and SOAP APIs on an access node 502. Access node 502 includes storage logic that is designed to effect storage of a file's data stream in blob storage node 506, and its associated metadata in metadata node 504, as described in more detail below.
A storage client first sends, at 1, a write request for F to an access node 502. The access node 502 sends, at 2, the data stream for F (Fd) to the scalable blob store 506 using a put operation. Fd is stored, at 3, on at least one node of the scalable blob store 506. Fd may be stored as an NTFS file, or as a chunk in some larger file that groups together several blobs. Fd may be replicated, at 4, within the blob store 506 prior to completing the put operation to guarantee durability of the put.
The blob store 506 returns, at 5, a locator token for Fd, Lt(Fd). The locator token can later be used as an argument to the blob store's get operation to locate and retrieve the contents of Fd.
The access node 502 sends, at 6, a metadata write request to the metadata store 504 to store both the file system metadata Fm, and locator token for F. Fm and Lt(Fd) are then stored, at 7, in the metadata store 504. SQL replication may be used, at 8, to replicate this information to guarantee its durability.
For retrieving file F, the above process is followed in reverse. That is, the locator token is first retrieved by an access node 502 from the metadata store 504 and then used to locate the appropriate blob in blob store 506.
A file replication service on nodes with associated file systems will have to check conditions 1, 2, and 3 described above. However, a replication service on a file system-less proxy does not need to take these conditions into account as it only stores objects in an object store that does not require file system constraints. This provides increased scalability and delayed conflict resolution.
Exemplary Embodiment—Sky Drive Architecture
In this particular embodiment, architecture 700 includes, by way of example and not limitation, consumer-side components and service side components which are delineated by the dashed line.
The consumer-side components include one or more replication clients 702 and one or more web clients 704. The service side components include one or more replication front ends 706 each of which includes a replication engine, one or more replication group redirection services 708, one or more web servers 710 that include a presentation layer, one or more storage access nodes 712, one or more metadata nodes 714 and one or more blob storage nodes 716. Storage access nodes include, in accordance with one embodiment, HTTP and SOAP APIs as mentioned above, along with replication business logic and storage logic which are described in more detail below.
In this embodiment, the file system metadata stored in the scalable file system's metadata nodes is augmented with the replicator's metadata described in Table 1. Updates to the unified metadata are made by replication business logic running in the access node 712 for the scalable file service so that both pieces of metadata may be transacted together. Data streams for files are stored in the scalable blob store 716 in their staged format, as described in
Consumer replication clients 702 perform state based replication with an instance of a state based replication engine hosted on replication front end 706. Consumers' replication clients participate in bidirectional replication with the replication service using the same replication protocols utilized when replicating with another consumer based replicator running on top of a regular file system, thus making replicating with the replication service backed by a blob store functionally equivalent to replication using a traditional file system.
Consumers are directed to a specific instance of a replication front end 706 via a replication group redirection service 708, thus ensuring that all replication for a set of files is performed on a single instance of the state based replication engine regardless of the number of consumers actively performing replication on that set of files. The affinity between file sets and a particular instance of a replication engine is utilized to enable in-memory locking and ordering of updates on the replication front ends 706, in order to reduce the number of update conflicts. Typically, the number of consumers replicating a given set of files will be small, i.e. not exceed a few hundred.
Web based clients 704 communicate with a presentation layer hosted on a web server 710 that renders the file system in HTML. The presentation layer translates the HTTP requests from the web based clients 704 and forwards these translated requests to scalable file service access node 712. The replication business logic embedded in the scalable file service access node 712 marshals files into a staged file format (e.g., as described in
In this example, the replication client 702 contacts, at 1, the replication group redirection service 708 to determine which replication front end to use to replicate file F. The replication client 702 then marshals file F, at 2, to produce a staged file, represented as Staged(F). One example of how this can be done is described in connection with
The blob store returns, at 8, a locator token for Staged(F), Lt(Staged(F)). The replication front end 706 then sends a metadata write request, at 9, to the storage access node 712 to write the replication metadata (the ID_Table data in
In one implementation, the consumer replication client 702 may upload the entire file F during step 3. In another implementation, file F may be streamed from the consumer replication client 702 to the scalable blob store 716, thus performing steps 3 through 5 as a pipelined operation. If an error occurs during steps 3 through 5, including, but not limited to: network disconnection, replication client failure, or a failure within the service, the replication front end 706 may attempt to store the information necessary to resume the transfer of Staged(F) with the consumer replication client at a later time. This information includes storing the partially transferred version of Staged(F) in the blob storage 716 and metadata about the partially transferred version of Staged(F) in the metadata store 714, including the client's identity and any other file version information required to resume the replication process.
In this example, the web client 704 uploads, at 1, file F to a web server 710. Responsively, web server 710 writes, at 2, file F to a storage access node 712. The replication business logic on the storage access node 712 produces, at 3, a staged version of F, Staged(F), including generating ID Table information, similar to the manner described in
The replication logic layer in the storage access node 712 sends a metadata update request, at 8, to the metadata store 714 containing the file system metadata Fm for F, the replication metadata (the ID_Table data in
If the metadata update in step 8 fails due to a transactional conflict between updates made by another web client or a consumer replication client, the replication logic layer in the storage access node retries the metadata update.
In one implementation, the web client 704 may upload the entire file F during step 1. In an another implementation, file F may be streamed from the web client 704 to the scalable blob store 716, thus performing steps 1 through 5 as a pipelined operation.
In this example, the web client 704 requests, at 1, file F from a web server 710. The web server 710 requests, at 2, file F to be read from storage access node 712. The storage logic on the storage access node 712 performs a lookup, at 3, of file F to obtain Lt(Staged(F)) using the metadata store 714. The metadata store 714 reads, at 4, the metadata (Fm, Lt(Staged(F), ID table data) for F from the SQL database to map F to Lt(Staged(F)). The metadata store 714 returns, at 5, Lt(Staged(F)) to the storage access node 712. The storage logic on the storage access node 712 uses Lt(Staged(F)) and issues a get request, at c6, for Staged(F) from the blob store 716. The blob store 716 reads, at 7, Staged(F) from storage and returns, at 8, Staged(F) to the storage access node 712.
The replication business logic on the storage access node 712 unmarshals, at 9, Staged(F), thus producing F. The storage access node 712 then returns file F, at 10, to the web server 710 which, in turn, returns file F, at 11, to the web client 704.
In an alternative implementation of
As noted above, a replication entity in the form of a replication hub can be used to implement the functionality described above. That is, one or more always-on replication hubs may be deployed to provide store-and-forward replication and higher network bandwidth than typical replication clients. Such hubs would typically be deployed in commercial data centers with significantly higher network connectivity than the typical replication node. The replication hubs can either be file-system less, thus providing replication of only the data streams for files, or replicate all of the file system metadata as described in the architecture section above.
The various embodiments described above provide a replication entity which implements a highly scalable file replication system. In at least some embodiments, a replication service provides a “drive in the sky” facility that can be used by individuals, such as subscribers, to synchronize their individual machines, such that their files are automatically replicated to a safe and always-on location. Alternatively or additionally, individuals such as subscribers can also access their files via a web-based interface when they are away from their machines. The inventive embodiments address and overcome at least some problems associated with scaling replication services to a very large number of subscribers. In at least some embodiments, these problems are addressed by virtue of a scalable, self-managing, and reliable binary large object (blob) store, as described above. Additionally, in at least some embodiments, the replication metadata is merged with scalable store metadata. Thus, any updates to the blob store (whether originating from replication with one of the subscriber machines, or due to changes made via the web interface) can transact both the blob store metadata and the replication meta-data at the same time. This mitigates synchronization issues between the replication metadata and the blob store and can thus obviate the need for rescans.
Although the invention has been described in language specific to structural features and/or methodological steps, it is to be understood that the invention defined in the appended claims is not necessarily limited to the specific features or steps described. Rather, the specific features and steps are disclosed as preferred forms of implementing the claimed invention.
|Cited Patent||Filing date||Publication date||Applicant||Title|
|US4934823 *||Nov 6, 1987||Jun 19, 1990||Hitachi, Ltd.||Staging method and system in electronic file apparatus|
|US5566331 *||Jan 24, 1994||Oct 15, 1996||University Corporation For Atmospheric Research||Mass storage system for file-systems|
|US5953729 *||Dec 23, 1997||Sep 14, 1999||Microsoft Corporation||Using sparse file technology to stage data that will then be stored in remote storage|
|US6460055 *||Dec 16, 1999||Oct 1, 2002||Livevault Corporation||Systems and methods for backing up data files|
|US6553391||Jun 8, 2000||Apr 22, 2003||International Business Machines Corporation||System and method for replicating external files and database metadata pertaining thereto|
|US6785882 *||May 24, 1999||Aug 31, 2004||Unisys Corporation||Process-driven tool interface for an object management system|
|US6874001 *||Oct 5, 2001||Mar 29, 2005||International Business Machines Corporation||Method of maintaining data consistency in a loose transaction model|
|US20010018686 *||Mar 14, 2001||Aug 30, 2001||Hitachi, Ltd.||Database processing method and apparatus using handle|
|US20020099728||Jan 19, 2001||Jul 25, 2002||Lees William B.||Linked value replication|
|US20020143798 *||Apr 1, 2002||Oct 3, 2002||Akamai Technologies, Inc.||Highly available distributed storage system for internet content with storage site redirection|
|US20020143888 *||Apr 1, 2002||Oct 3, 2002||Akamai Technologies, Inc.||Scalable, high performance and highly available distributed storage system for internet content|
|US20020147774 *||Apr 1, 2002||Oct 10, 2002||Akamai Technologies, Inc.||Content storage and replication in a managed internet content storage environment|
|US20020174103||Apr 30, 2002||Nov 21, 2002||International Business Machines Corporation||System and method for updating external file referenced by database with transactional consistency using SQL|
|US20030037029 *||Aug 15, 2001||Feb 20, 2003||Iti, Inc.||Synchronization of plural databases in a database replication system|
|US20030182319 *||Mar 25, 2002||Sep 25, 2003||Michael Morrison||Method and system for detecting conflicts in replicated data in a database network|
|US20040267843 *||Jun 26, 2003||Dec 30, 2004||International Business Machines Corporation||Replication of binary large object data|
|US20050027757 *||Aug 1, 2003||Feb 3, 2005||Rick Kiessig||System and method for managing versions|
|US20050071389 *||Sep 29, 2003||Mar 31, 2005||International Business Machines Corporation||High availability data replication of smart large objects|
|US20050091287 *||Nov 1, 2004||Apr 28, 2005||Eric Sedlar||Database-managed file system|
|US20050160120||Jan 21, 2004||Jul 21, 2005||Todd Stephen J.||Methods and apparatus for modifying a retention period for data in a storage system|
|US20050203962 *||Mar 9, 2004||Sep 15, 2005||Dong Zhou||Framework and associated apparatus for the adaptive replication of applications with server side code units|
|US20050262173 *||Jul 20, 2005||Nov 24, 2005||Microsoft Corporation||System and method providing virtual applications architecture|
|US20060112244 *||Nov 24, 2004||May 25, 2006||International Business Machines Corporation||Automatically managing the state of replicated data of a computing environment|
|US20060179082 *||Feb 9, 2005||Aug 10, 2006||Boyd Kenneth W||Method, system and article of manufacture for metadata replication and restoration|
|US20060235893 *||Apr 15, 2005||Oct 19, 2006||Emc Corporation||Methods and apparatus for managing the storage of content|
|US20060248038 *||Apr 29, 2005||Nov 2, 2006||Marc Kaplan||System and method of handling file metadata|
|US20060294164 *||Jun 23, 2005||Dec 28, 2006||Emc Corporation||Methods and apparatus for managing the storage of content in a file system|
|EP1462960A2||Mar 18, 2004||Sep 29, 2004||Microsoft Corporation||Consistency unit replication in application-defined systems|
|1||*||Alex Sim, Junmin Gu, Arie Shoshani, and Vijaya Natarajan, DataMover: Robust terabyte-scale multi-file replication over wide-area networks, Lawrence Berkeley National Laboratory, University of California, 2004, Paper LBNL' 54854.|
|2||*||Alex Sim, Junmin Gu, Arie Shoshani, and Vijaya Natarajan, DataMover: Robust terabyte-scale multi-file replication over wide-area networks, Lawrence Berkeley National Laboratory, University of California, 2004, Paper LBNL′ 54854.|
|3||Extended European Search Report, mailed on Dec. 29, 2009, from application No. PCT/US2006049663, 8 pages.|
|4||Windows Server 2003 R2: Extending Support for Branch Offices, May 2005, pp. 1, 5.|
|Citing Patent||Filing date||Publication date||Applicant||Title|
|US7992037 *||Jul 29, 2009||Aug 2, 2011||Nec Laboratories America, Inc.||Scalable secondary storage systems and methods|
|US8442983 *||May 14, 2013||Commvault Systems, Inc.||Asynchronous methods of data classification using change journals and other data structures|
|US9078144 *||May 2, 2012||Jul 7, 2015||Nokia Solutions And Networks Oy||Signature enabler for multi-vendor SON coordination|
|US9146937||Jun 22, 2012||Sep 29, 2015||Cohort Fs, Llc||Client-based data replication|
|US9213699 *||Apr 30, 2008||Dec 15, 2015||Aspect Software Inc.||Super-object in administering system|
|US9253666 *||Jun 26, 2015||Feb 2, 2016||Nokia Solutions And Networks Oy||Signature enabler for multi-vendor SON coordination|
|US20090276452 *||Nov 5, 2009||Aspect Software, Inc.||Super-object in administering system|
|US20100064166 *||Mar 11, 2010||Nec Laboratories America, Inc.||Scalable secondary storage systems and methods|
|US20110161327 *||Dec 23, 2010||Jun 30, 2011||Pawar Rahul S||Asynchronous methods of data classification using change journals and other data structures|
|US20140040196 *||Aug 2, 2013||Feb 6, 2014||Egnyte, Inc.||System and Method for Event-Based Synchronization of Remote and Local File Systems|
|US20140040197 *||Aug 2, 2013||Feb 6, 2014||Egnyte, Inc.||System and Method for Event-Based Synchronization of Remote and Local File Systems|
|U.S. Classification||707/612, 714/15, 707/614, 714/6.11|
|International Classification||G06F17/30, G06F7/00|
|Mar 31, 2006||AS||Assignment|
Owner name: MICROSOFT CORPORATION, WASHINGTON
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNORS:TEODOSIU, DAN;STEERE, DAVID C.;ANNAMALAI, MUTHUKARUPPAN;AND OTHERS;REEL/FRAME:017413/0216;SIGNING DATES FROM 20060227 TO 20060310
Owner name: MICROSOFT CORPORATION,WASHINGTON
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNORS:TEODOSIU, DAN;STEERE, DAVID C.;ANNAMALAI, MUTHUKARUPPAN;AND OTHERS;SIGNING DATES FROM 20060227 TO 20060310;REEL/FRAME:017413/0216
|Nov 26, 2013||FPAY||Fee payment|
Year of fee payment: 4
|Dec 9, 2014||AS||Assignment|
Owner name: MICROSOFT TECHNOLOGY LICENSING, LLC, WASHINGTON
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:MICROSOFT CORPORATION;REEL/FRAME:034543/0001
Effective date: 20141014